Classification Of Polarimetric SAR Images Based On Selective Ensemble Learning

The system of polarimetric synthetic aperture radar technology(Pol SAR) takes advantage the differences of the channels’ relative phase information to get the quantitative reflection comparing with the general SAR. The advantage has a great value in the fields of target detection and identification, precision agriculture, disaster monitoring. Therefore, the study of Pol SAR also has got a widespread attention in the scientific community. For the traditional image classification, a single classifier often does not cons ider all problems so that there is a big difference with the actual situation. To solve this problem, ensemble learning is proposed, which trains a number of different learning machines to get the final results by integrating combination so as to get more ideal and better results than the single learner. But it does not mean, the larger number of learners are used, the better results can be get. The use of a large number of learners may lead to the negative effects. To solve this problem, the concept of selective ensemble learning is proposed, i.e., when a group of individual learners are available, only selecting some of them to integrate can obtain a better result than integrating all the learners. Combined with the characteristics of polarimetric SAR image, this paper proposed selective ensemble algorithm for Pol SAR image classification to improve the effect of classification and promote the development of polarimetric SAR image field. This paper mainly includes the following three aspects:(1) The basic scattering model of polarimetric SAR images and representation of their properties are studied. The polarization characteristics and the non-polarization feature extraction methods of the polarization SAR image are also studied. Furthermore, this paper studies the extraction methods based on the target polarization scattering model decomposition and texture fea tures.(2) The basic theory of ensemble learning and selective ensemble learning are researched. Combined with the characteristics of polarimetric SAR images, this paper studies the construction of base classifiers of selective ensemble, i.e.SVM and NN. In this paper, the data of EMISAR in Foulum and ESAR in Oberpfaffenhofen are used.(3) According to the characteristics of the selective ensemble, two kinds of selective ensemble algorithms are proposed, which are selective ensemble based on Fuzzy C-means Clustering and selective ensemble based on GA algorithmCombined with the characteristics of the two algorithms, the two algorithms are applied to polarimetric SAR image classification to obtain the accuracy of full images and each type of test samples. Comparing the classification results of selective ensemble and single classifier, assess the advantages of superiority of algorithms.